On scientific understanding with artificial intelligence

Mario Krenn,1, 2, 3, 4, ∗ Robert Pollice,2, 3 Si Yue Guo,2 Matteo Aldeghi,2, 3, 4 Alba

Cervera-Lierta,2, 3 Pascal Friederich,2, 3, 5 Gabriel dos Passos Gomes,2, 3 Florian Häse,2, 3, 4, 6
Adrian Jinich,7 AkshatKumar Nigam,2, 3 Zhenpeng Yao,2, 8, 9, 10 and Alán Aspuru-Guzik2, 3, 4, 11, †
1
Max Planck Institute for the Science of Light (MPL), Erlangen, Germany.
2
Chemical Physics Theory Group, Department of Chemistry, University of Toronto, Canada.
3
Department of Computer Science, University of Toronto, Canada.
4
Vector Institute for Artificial Intelligence, Toronto, Canada.
5
Institute of Nanotechnology, Karlsruhe Institute of Technology, Eggenstein-Leopoldshafen, Germany.
6
Department of Chemistry and Chemical Biology, Harvard University, Cambridge, USA.
7
Division of Infectious Diseases, Weill Department of Medicine,
Weill-Cornell Medical College, New York, USA.
8
Center of Hydrogen Science, Shanghai Jiao Tong University,
800 Dongchuan Road, Shanghai 200240, China.
arXiv:2204.01467v1 [cs.CY] 4 Apr 2022

9
The State Key Laboratory of Metal Matrix Composites,
School of Materials Science and Engineering, Shanghai Jiao Tong University,
800 Dongchuan Road, Shanghai 200240, China.
10
Innovation Center for Future Materials, Zhangjiang Institute for Advanced Study,
Shanghai Jiao Tong University, 429 Zhangheng Road, Shanghai 201203, China.
11
Canadian Institute for Advanced Research (CIFAR) Lebovic Fellow, Toronto, Canada.
Imagine an oracle that correctly predicts the outcome of every particle physics experiment, the
products of every chemical reaction, or the function of every protein. Such an oracle would revolu-
tionize science and technology as we know them. However, as scientists, we would not be satisfied
with the oracle itself. We want more. We want to comprehend how the oracle conceived these
predictions. This feat, denoted as scientific understanding, has frequently been recognized as the
essential aim of science. Now, the ever-growing power of computers and artificial intelligence poses
one ultimate question: How can advanced artificial systems contribute to scientific understanding
or achieve it autonomously?
We are convinced that this is not a mere technical question but lies at the core of science. There-
fore, here we set out to answer where we are and where we can go from here. We first seek advice
from the philosophy of science to understand scientific understanding. Then we review the current
state of the art, both from literature and by collecting dozens of anecdotes from scientists about how
they acquired new conceptual understanding with the help of computers. Those combined insights
help us to define three dimensions of android-assisted scientific understanding: The android as a I)
computational microscope, II) resource of inspiration and the ultimate, not yet existent III) agent
of understanding. For each dimension, we explain new avenues to push beyond the status quo and
unleash the full power of artificial intelligence’s contribution to the central aim of science. We hope
our perspective inspires and focuses research towards androids that get new scientific understanding
and ultimately bring us closer to true artificial scientists.

I. INTRODUCTION distinguish between this and communicating with a

Artificial Intelligence (A.I.) has recently been called
a “new tool in the box for scientists”[1] and that “ma-
chine learning with artificial networks is revolutioniz-
ing science“[2]. Additionally, it has been conjectured
“that machines could have a significantly more cre-
ative role in future research.” [3]. For instance, it has
even been postulated that “[t]he new goal of theoret-
ical chemistry should be that of providing access to
a chemical ’oracle’: an A.I. environment which can
help humans solve problems, associated with the fun-
damental chemical questions of the fourth industrial
revolution [...], in a way such that the human cannot
∗ mario.krenn@mpl.mpg.de
† alan@aspuru.com
human expert” [4].
However, this excitement has not been shared
among all scientists. Specifically, it has been ques-
tioned whether advanced computational approaches
can go beyond numerics [5–9] and contribute funda-
mentally to one of the essential aims of science, that
is, gaining of new scientific understanding [10–12].
In this work, we address how artificial systems can
contribute to scientific understanding – specifically,
what is the state-of-the-art and how we can push fur-
ther. Besides a thorough literature review, we sur-
veyed dozens of scientists at the interface of biol-
ogy, chemistry or physics on the one hand, and arti-
ficial intelligence and advanced computational meth-
ods. These personal narratives focus on the concrete
discovery process of ideas and are a vital augmenta-
tion to the scientific literature. We put the literature
and personal accounts in the context of a philosophi-
 2

Figure 1. How can Androids contribute to new scientific understanding? In addition to scientific literature,
we take inspiration from the philosophy of science and from dozens of stories provided by active computational natural
scientists. Thereby we identify three fundamental dimensions of computer-assisted scientific understanding. From there,
we look into the future and develop a roadmap on how to develop Androids that can contribute to understanding – the
essential aim of science.

cal theory of Scientific Understanding recently devel-
oped by Dennis Dieks and Henk de Regt [12, 13], who
was awarded the Lakatos Award in 2019 for the de-
velopment of this theory. We thereby introduce three
fundamental dimensions for scientific androids1 con-
tribution towards new scientific understanding:
I) Androids acting as a microscope in the re-
sponses, i.e., akin to an instrument revealing
properties of a physical system that are other-
wise difficult or even impossible to probe. Hu-
mans then lift these insights to scientific under-
standing.
II) Androids acting as muses, i.e., sources of inspi-
ration for new concepts and ideas that are sub-
sequently understood and generalized by human
scientists.
1 We encapsulate all advanced artificial computational systems
under androids, independent of their working principles. In
this way, we are focusing on the operational objective rather
than the methodology.
III) Lastly, in an ultimate dimension of android-
assisted scientific understanding, computers are
the agents of understanding. While we have not
found any evidence of computers acting as true
agents of understanding in science yet, we out-
line important characteristics of such an artifi-
cial system of the future and potential ways to
achieve it.
In the first two dimensions, the android enables hu-
mans to gain new scientific understanding while in the
last one the machine gains understanding itself. These
classes enable us to layout a vibrant and mostly unex-
plored field of research, which will hopefully manifest
itself as a guiding star for future developments of ar-
tificial intelligence in the natural sciences.
The goal of this perspective is to put Scientific
Understanding back to the limelight – where we are
convinced it belongs. We hope to inspire physicists,
chemists and biologists and A.I. researchers to go be-
yond the status quo, focus on these central aims of
science, and revolutionize computer-assisted scientific
understanding. In that way, we believe that androids

will become true agents of understanding that con-
tribute to science in a fundamental and creative way.
II. SCIENTIFIC UNDERSTANDING
Let us imagine an oracle providing non-trivial pre-
dictions that are always true. While such a hypo-
thetical system would have a very significant scien-
tific impact, scientists would not be satisfied. We
want “to be able to grasp how the predictions are gen-
erated, and to develop a feeling for the consequences
in concrete situations” [13]. Colloquially, we refer to
this goal as “understanding” – But what does that
really mean? Can we find criteria for scientific under-
standing? To do that, we seek guidance from the field
of philosophy of science. Notably, while hardly any
scientist would argue against “understanding” as an
essential aim of science (next to explanation, descrip-
tion and prediction [14]), this view was not always
accepted by philosophers. Specifically, Carl Hempel,
who made foundational contributions clarifying the
meaning of scientific explanation, argued that “un-
derstanding” is subjective and merely a psychologi-
cal by-product of scientific activity and is therefore
not relevant for the philosophy of science [15]. Other
philosophers criticized these rather unsatisfying con-
clusions, and they tried to formalize what scientific
understanding means. Proposals include that under-
standing is connected to the ability to build causal
models (Lord Kelvin said “It seems to me that the
test of ’Do we or not understand a particular subject
in physics?’ is, ’Can we make a mechanical model
of it?’ ”[13]), connected to providing visualizations
(or Anschaulichkeit, as its strong proponent Erwin
Schrödinger called it[16, 17]) or that understanding
corresponds to providing unification [18, 19].
In recent years, Henk de Regt and Dennis Dieks
have developed a new theory of scientific understand-
ing, which is both contextual and pragmatic [12–14].
Importantly, they find that techniques such as visu-
alization or unification are “tools for understanding”,
thereby unifying previous ideas in one general frame-
work. Their theory is agnostic to the specific “tool”
being used, making it particularly useful for applica-
tion in scientific disciplines. They extend crucial in-
sights by Werner Heisenberg [20] and rather than in-
troducing mere theoretical or hypothetical ideas, the
main motivation behind their theory is that a “satis-
factory conception of scientific understanding should
reflect the actual (contemporary and historical) prac-
tice of Science”. Put simply, they argue that:
A phenomenon P can be understood if there
exists an intelligible theory T of P such that
scientists can recognise qualitatively charac-
teristic consequences of T without performing
exact calculations [12, 13].
3
Concretely, de Regt and Dieks define two inter-
linked criteria:
1. Criterion of Understanding Phenomena:
A phenomenon P can be understood if a theory
T of P exists that is intelligible.
2. Criterion for the Intelligibility of Theo-
ries: A scientific theory T is intelligible for sci-
entists (in context C) if they can recognise qual-
itatively characteristic consequences of T with-
out performing exact calculations.
We decided to use this specific theory because of
one particular strength: We can use it experimen-
tally to evaluate whether scientists have understood
new concepts or ideas, rather than by inspecting their
methodology, by simply looking at the scientific out-
come and the consequences. This also coincides with
Angelika Potochnik’s argument that “understanding
requires successful mastery, in some sense, of the tar-
get of understanding”[11]. We will follow this ap-
proach and, consequently, here explore its relationship
to the role of A.I. in science. Accordingly, we believe
we can significantly advance A.I.’s contribution to this
central aim of Science if we have a clear picture of
how scientists gain conceptual understanding, and in-
stil it to artificial systems afterwards. We approach
this goal by applying ideas of de Regt and Dieks di-
rectly to android assisted science (and ultimately, to
android scientists themselves).
III. WHAT IS NEXT?
A. Beyond Re-Discovery
In recent years, scientists at the interface between
A.I. and the natural sciences tried to rediscover scien-
tific laws or concepts with machines. The question is,
however, whether an android is capable of contribut-
ing to new scientific understanding if it can rediscovers
physical laws and concepts, such as the heliocentric
world view [21], the arrow of time [22] or mechani-
cal equations of motions [23]? We believe that this
is not guaranteed. The human creators of these an-
droids know what they are looking for in these case
studies. Therefore, it is unclear how both conscious
and unconscious biases (in the broadest sense, e.g., by
choosing particular representations) in the code or the
data analysis can be prevented. Consequently, even if
an algorithm can rediscover interesting physical phe-
nomena, we cannot know whether and how they can
be used to advance Science by helping to uncover new
scientific understanding.
Hence, we believe we need to go beyond rediscovery
tasks. Therefore we focus explicitly on the question
of how to get new scientific understanding.

B. Beyond Discovery
Importantly, other central aims of science such as
prediction and discovery can lead to scientific and
technological disruptions while not directly contribut-
ing to scientific understanding as discussed above
[11, 14]. For instance, imagine the hypothetical dis-
covery of the hitherto best material for energy stor-
age that could revolutionize batteries. However, this
game-changing discovery would not qualify as under-
standing if chemists could not use the underlying prin-
ciples fruitfully in other contexts (without computa-
tion).
Similarly, the recent breakthrough in protein fold-
ing will undoubtedly change the landscape of biochem-
istry. However, so far, AlphaFold is a black box – an
oracle[24, 25]. As such it does not directly provide
new scientific understanding in the sense of de Regt
and Dieks (but could of course in the future enable
humans to gain new scientific understanding). Hence,
we believe we must go beyond artificial discoveries in
science.
C. Where to go from here?
The ultimate goal is to get new understanding from
androids. Loosely speaking, we want to find new ideas
or concepts that we can apply and use in different
situations without (complete) computations.
This article aims to explain precisely what such
a goal requires, what previous approaches have
achieved, and how we can go further. We want to
clearly lay out this underappreciated but essential re-
search question and thereby give a clear goal for the
future of A.I. in the natural sciences.
IV. THREE DIMENSIONS OF
COMPUTER-ASSISTED UNDERSTANDING
We use scientific literature and personal anecdotes
of dozens of scientists, and the context of the philos-
ophy of science, to introduce a new classification of
androids contribution to scientific understanding2 . It
helps to see diverse unexplored journeys that can be
investigated in the future.
An android can act
I) as a computational microscope, providing in-
formation not (yet) attainable by experiment
II) as a resource of inspiration or an artificial
muse, expanding the scope of human imagina-
tion and creativity.
2 Wecall the classification dimensions, as they are independent
and non-exclusive.
4
In those two classes, the human scientist is essential
to take the new insight and inspiration and develop it
to full understanding. Finally, an android can be
III) an agent of understanding, replacing the hu-
man in generalizing observations and transfer-
ring scientific concepts to new phenomena.
We stress that these three classes should not be un-
derstood dogmatically but rather guide future possi-
bilities. In the following sections, based on concrete
examples, we discuss each class in more detail and pro-
pose avenues for pushing the boundaries of the current
computational faculties.
A. Computational microscope for scientific
understanding
Microscopes are devices that enable us to inves-
tigate objects and phenomena imperceptible to the
naked eye. In a similar way, computational micro-
scopes enable the investigation of objects or processes
that we cannot visualize or probe in any other way.
One main objective is to simulate biological, chemical
or physical processes that happen at length and time
scales not perceivable by experiment.
As we are interested in understanding, the new
computer-generated data needs to be generalized to
other contexts without complete computation[13]. We
show now two concrete examples.
The first example is molecular dynamics simulations
of the SARS-CoV-2. The authors uncovered new bio-
logical functions that show different behaviours in the
open and closed conformations of the spike protein.
This explanation changed the view upon glycans in
biological systems and inspired new ways to analyze
these systems without the need to perform full com-
putations [26].
In the second example, the authors describe how
molecular dynamics simulations helped to uncover
fundamental patterns called glycoblocks. The sys-
tematic use of glycoblocks can both be used to un-
derstand sequence-structure-property relationships of
biomolecules and can also inform the design of syn-
thetic structures with desired functions without the
need for simulating the entire system [27].
The next computational microscope
A computational microscope aims to provide data
via computation that are not (yet) accessible by ex-
periments that humans can understand. How could
we make computational microscopes even more in-
sightful and make it easier for human scientists to use
this data to gain scientific understanding? There are
two vibrant directions going forwards. First, more ad-
vanced computational systems will allow to analyze of
 5

Figure 2. The future computational microscope. We envision two types of advances in the next-generation
computational microscopes which aim to advance genuine scientific understanding. First (left), larger and more complex
computations will allow the computational observation of phenomena not accessible so far. There, new computational
paradigms will play a significant role, such as Graphical Processor Units (GPU), Tensor Processor Units (TPU), Optical
Processor Units (OPU) and – ultimately – quantum computers. Second (right), new ways to represent the highly complex
data will advance our ability to sense structure and recognize underlying patterns. The involvement of all our senses
could, for sensing computer-generated data, be an exciting pathway to advanced understanding.

more complex physical systems. Second, representing
the information in a more interpretable way will help
to lift the indications from computers to true scientific
understanding.
More Complex Systems – One obvious but nev-
ertheless important research direction is increasing
the complexity as well as the accuracy of computer
simulations[28]. For example, increasing the size of
the systems, the time-scale of the simulations, the
number of interactions that can be modelled will sig-
nificantly increase the applicability in complex dy-
namic systems. In general, this can be achieved
by either algorithmic improvements or hardware im-
provements, or both. In that regard, we expect that
modern neural network technologies together with ad-
vanced hardware such as GPUs, TPUs or even OPUs
[29, 30] will have an enormous impact. Furthermore,
the recent progress in experimental quantum com-
puting for quantum chemistry [31] and physics[32–34]
promises that entirely new algorithms, based on quan-
tum mechanics itself, will play an important role in
this area [35, 36]. Algorithmic improvements could in-
volve adaptive and intelligent resolution during simu-
lation and advanced visualization methods[13], which
directly leads to the second future techniques:
Full spectrum of senses – We believe that human
scientists can get more out of data if the full capabil-
ities of all our senses are addressed. At the moment,
we analyze data largely in (potentially animated) 2-
dimensional pictures. As a first step, we believe that
real 3D environments (realized either via virtual or
augmented reality glasses, or holography) will signifi-
cantly help in understanding complex systems or com-
plex data. Initial advances in that regard have been
demonstrated in the domain of chemistry [37–39], and
we expect this to become a standard tool for scientists
to advance scientific understanding. In addition, we
expect that going beyond the visual sense can open
entirely new ways to experience scientific data. For
example, the auditory sense is excellent in detecting
structure or symmetries in (periodic) time-dependent
data [40]. Furthermore, including the sense of touch,
smell and taste could further expand the horizon of
experiences. We expect that in order to realize that,
physical scientists need to work closely together with
psychologists and neurologists (and potentially even
with artists), to develop suitable data representations
that can be efficiently recognized by scientists with
all their senses. An ultimate, admittedly futuristic
version of a computational microscope could circum-
vent the receptors of human senses and instead use
a computer-brain interface to enhance further experi-
encing computed data.
B. Resource of inspiration for scientific
understanding
Surprising and creative ideas are the foundation of
Science. Computer algorithms are a means to provoke
such ideas systematically, thereby significantly accel-
erating scientific and technological progress. Already
70 years ago, Alan Turing realized that computers
could surprise their human creators: “Machines take
me by surprise with great frequency. This is largely be-
cause I do not do sufficient calculation to decide what
to expect them to do, or rather because, although I
do a calculation, I do it in a hurried, slipshod fash-
ion, taking risks.” and ”Naturally I am often wrong,
and the result is a surprise for me for by the time
the experiment is done these assumptions have been
forgotten“ [41].
A much more recent study provides stories by

dozens of researchers of artificial life and evolution.
They demonstrate in an impressive way how computer
algorithms can surprise their human creators and lead
to behaviour that the authors would denote as cre-
ative[42]. Accordingly, we believe that androids can
be artificial muses of Science in a metaphorical sense.
Those examples demonstrate that computers can
indeed be used as a source of surprises. But what are
the most general ways to get inspirations from com-
puters? And how can they be lifted by humans to true
scientific understanding? We will outline a number of
ways to develop ways to provoke surprising behaviours
of algorithms and use their solutions, internal or exter-
nal states as a source of inspiration for new scientific
ideas.
The future resource of inspiration
Identifying surprises in data – Exceptional data
points or unexpected regularities obtained from ex-
periments or simulations can surprise human scien-
tists and inspire new ideas and concepts. Our survey
shows that these exceptional points are usually iden-
tified by humans, such as the following two examples,
which use high-throughput computations in chemistry
[43] and quantum optics [44, 45].
The first example deals with a surprising phase of
crystal structures in high-pressure physics. There, the
authors found an unexpected stable configurations of
alternating NH2 and NH4 layers, rather than a dense
NH3 phase. The authors conceptualized this phe-
nomenon as spontaneous ionization, a common pro-
cess in acid-base chemistry, which is now a widely
accepted phenomenon in the high-pressure phase di-
agram of NH3 . Spontaneous ionization in the high-
pressure behaviour of matter has become a more gen-
eral principle that can be used without performing any
simulations [46].
In the second example, a search for new quan-
tum experiments uncovered a solution with consid-
erable larger quantum entanglement than expected.
The authors understood the underlying principles and
thereby discovered a new concept of entanglement
generation [47, 48]. The principle can be used without
any computation and, for example, acts now as a new
representation in more advanced artificial intelligence
systems for quantum physics [49], demonstrating the
application of the computer-inspired idea in more gen-
eral different contexts.
In contrast to these examples and many others from
literature and from personal accounts, the anomalies
could manifest themselves in a more involved com-
bination of variables, which might be very difficult
for humans to grasp. Accordingly, applying advanced
statistical methods and machine learning algorithms
(e.g., see reference [50]) to this type of problem will
be an important future research direction. Exciting
6
works into the direction of autonomous anomaly de-
tection have been applied on scientific data from the
Large Hadron Collider (LHC) at CERN [51–53]. Such
techniques have the potential to identify new physics
signatures, which can then be conceptualized and un-
derstood by human physicists [54, 55]. Neural net-
works that autonomously discover symmetries could
become an efficient discovery tool for outliers in sci-
entific data where the underlying rules might not be
known beforehand [56, 57].
Estimating the confidence of predictions will be one
method to directly search for anomalies in data [58].
The ability to uncover hidden regularities was very
recently demonstrated in mathematics, where an A.I.
hinted on relations between previously unconnected
invariants in knot theory, which allowed mathemati-
cians to conjecture and prove new theorems [59]. Al-
ternatively, an A.I. capable of constructing new scien-
tific hypotheses could uncover outliers or unexpected
patterns that are not discernible with standard statis-
tical methods.
It would be truly exciting to see an A.I. uncover hid-
den patterns or irregularities in scientific data previ-
ously overlooked by humans, which leads to new ideas
and, ultimately, to new conceptual understanding. As
of now, we are not aware of cases like that.
The data points for these systems could be ob-
tained from computational methods (involving those
described in section IV A), with exciting opportunities
for mathematics or theoretical physics [60]. Alterna-
tively, the data could be obtained directly from exper-
iments. Here we can imagine a closed-loop approach
where an algorithm tries to explore the environment
and steer the exploration into unexpected regions. If
the data-source is an experiment, this future system
will require access to complex lab automation with
large parameter spaces to explore, as demonstrated
recently in biology [61], chemistry [62–67] or physics
[68, 69].
Identifying surprises in the scientific literature –
The number of scientific papers in essentially every
scientific domain is growing enormously[70]. Conse-
quently, researchers have to specialise in narrow sub-
disciplines, which makes finding new interdisciplinary
ideas difficult. In the future, we believe that comput-
ers will be able to use the scientific literature in an
automated way[71–74] and identify exceptional and
surprising phenomena for further investigation. While
the large-scale automated analysis of the scientific lit-
erature, to our knowledge, has not yet been able to
induce new scientific understanding, there is signifi-
cant progress in the field. One promising approach
towards this goal is unsupervised word embedding of
a large corpus of scientific papers. In that technique,
the content of the scientific literature is transformed
into a high-dimensional vector space. Recently, this
technique has been applied in the domain of mate-
rial science [75] and rediscovered central scientific con-

Figure 3. The future re-source of inspiration: An android could act as a computational muse and inspire the
human scientist by (A) identifying surprises in data, (B) identifying surprises in the scientific literature, (C) finding
surprising concepts by inspecting models or (D) by probing the behaviour of artificial agents or (E) my finding new
concepts from interpretable solutions.
cepts such as the periodic table of the elements. Ad-
ditionally, the results also suggested the existence of
previously undiscovered structure-property relation-
ships. Examples include new candidates for thermo-
electric materials. Moreover, several other advanced
computational techniques are being developed in ma-
terial science to extract knowledge from the scien-
tific literature and investigate it systematically by A.I.
technologies[76], and can lead to complex scientific
conclusions as demonstrated for instance on zeolite
transformations [77].
An alternative approach aims to build semantic
knowledge networks from large bodies of scientific lit-
erature. In these networks, scientific concepts are
nodes, and edges carry relational information. In the
simplest case, that means two scientific concepts are
mentioned in the same scientific paper [78, 79]. Thus,
scientific knowledge is represented as an evolving net-
work, which can be used to identify both islands and
unexplored regions of the scientific literature. This
type of network was used in biochemistry to identify
efficient global research strategies [78] and in quantum
physics to predict and suggest future research direc-
tions [79]. Advances in A.I. technology could improve
this type of system significantly. For example, natural
language processing architectures such as BERT [80],
or GPT3 [81] could help extract more scientific knowl-
edge from research papers, and large graph-based neu-
ral networks could improve the prediction of new re-
search topics from semantic networks [82].
Surprising concepts by inspecting models – We also
expect considerable progress by rationalising what
A.I. algorithms have learned in order to solve a specific
problem, i.e., explainable or interpretable A.I. [83–86].
One idea towards this goal is inspired by DeepDream-
ing, a method first used in computer vision [87, 88].
Put simply, the idea is to invert a neural network
7
and probe its behaviour. Recently, this approach has
been applied to rediscover thermodynamical proper-
ties [22], and design principles for functional molecules
[89]. An alternative and remarkable application is the
disentanglement of variables in neural networks [90].
The goal is to understand the internal representation
the neural network has learned. Recently, astronom-
ical data, represented in geocentric coordinates, was
used to train a neural network and disentanglement
of variables enabled the rediscovery of heliocentric co-
ordinates via the internal representation of the model
[21]. In a related study, using gradient boosting with
decision trees, feature importance has been used to
explain properties of molecules, and quantum optics
circuits [91]. Related to this is a study where the inter-
nal representation of an unsupervised deep generative
model for quantum experiments has been inspected
to understand the model’s internal worldview[92]. In
the chemical domain, counterfactual explanations for
machine learning models have been demonstrated to
produce rationale behind a model’s prediction. Coun-
terfactual explanations illustrate what differences to
an event or instance would generate a change in an
outcome. Wellawatte et al. [93] showed how this can
be achieved in a model-independent way (it has been
demonstrated for random forest, sequence models and
graph neural networks), indicating great future poten-
tials for opening the black-box of AI in science. Albeit
not in science, recent work has investigated what the
chess-playing A.I AlphaZero has learned about chess
and how human-like knowledge is encoded in the in-
ternal representation [94]. The concepts rediscovered
in all of those works were not new, and thus the most
important challenge for the future is to learn how to
extract previously unknown concepts. Progress to-
wards resolving that challenge will be essential in the
near future to inspire new scientific ideas.

New concepts from interpretable solutions – Rather
than getting inspirations from the A.I. algorithms
themselves, scientists can also be surprised by the cor-
responding solutions. When solutions are represented
in an interpretable way, they can provoke new ideas
and lead to new concepts. An example of interpretable
representations is a mathematical formula. Thus,
scientists can inspect formulae derived by computer
algorithms to solve mathematical problems directly
and derive more general solution strategies. Several
publications demonstrated extracting symbolic mod-
els from experimental data of mechanical systems
[23, 95], of quantum systems [96] and in astronomy
[97]. It will be exciting to see how these approaches,
e.g., combined with methods such as causal inference
[98], can be improved to propose reasonable physical
models of unknown systems that advance scientific un-
derstanding. Altogether, exciting advances have been
achieved in the field of mathematics[99, 100], and we
foresee similar approaches making a significant impact
in the physical sciences as well.
One concrete, recent example in astronomy is the
rediscovery of Newton’s law of gravitation from real-
world observational data of planets and moons in our
solar system from the last 30 years [101]. The appli-
cation of graph neural networks allowed for the high-
quality prediction of the object’s motion. Further-
more, a symbolic regression technique called PySR
(introduced in [97]) was able to extract reasonable
mathematical expressions for the learned behaviour.
Interestingly, besides the equations of motions, the
method simultaneously predicts the masses of the
planetary objects correctly. The technique required
the assumption of several symmetries and other phys-
ical laws. It will be interesting to see whether these
prerequisites can be reduced further and how related
approaches can be applied to modern physics ques-
tions.
Another concrete example of this methodology has
been showcased in the field of quantum optics [49].
There, an A.I. algorithm with a graph-theoretical rep-
resentation of quantum optical setups designs config-
urations for previously unknown quantum systems.
The final solutions were represented in a physically-
interpretable graph-theoretical representation. From
there, human scientists can quickly interpret the un-
derlying reasons why the solutions work and apply
it in other contexts without further computation.
Accordingly, developing interpretable representations
and methods to extract underlying concepts in other
domains will be an important future research direc-
tion.
Probing the behaviour of artificial agents – Another
only rarely explored opportunity is interpreting the
behaviour of machines when tasked to solve a scien-
tific problem [102]. Algorithms that take actions such
as genetic algorithms or reinforcement learning agents
adopt policies to navigate the problem space. Human
8
scientists can observe how they navigate this space.
Instead of following a strict external reward, e.g., max-
imise a specific property of a physical system, intrinsic
rewards such as artificial curiosity can be implemented
[103, 104]. Instead of maximizing directly some func-
tions, the artificial agent tries to learn and predict
the behaviour of the environment. It then chooses
actions that lead to situations it cannot predict well,
thus maximizing its own understanding of the envi-
ronment. It has been shown using curious agents in
simulated virtual universes [105] and robot agents in
real laboratories [67] that curiosity is an efficient ex-
ploration strategy. Alternative intrinsic rewards for
artificial agents are computational creativity[106, 107]
and surprise [108]. These intrinsic rewards can pro-
duce exceptional and unexpected solutions, ultimately
inspiring human scientists.
C. Agent of Understanding
The third and final class we consider are algorithms
that can autonomously acquire new scientific under-
standing, a feat that has neither been described by
the respondents of our survey nor in the scientific lit-
erature. Therefore, we will approach this class by list-
ing the requirements of these agents, proposing tests
to detect their successful realization and speculating
what such computer programs could look like.
First, it is important to realize that finding new
scientific understanding is context-dependent. What
is new depends on whether we consider an individ-
ual scientist and their field of expertise, a scientific
domain, the whole scientific community or even the
entire scientific endeavour throughout history. Hence,
true agents of understanding must be able to evaluate
whether an insight is new, at least in the context of
a specific scientific domain that requires access to the
knowledge of a scientific field.
Secondly, de Regt emphasized the importance of
underlying scientific theories that allow us to recog-
nize qualitatively characteristic consequences [12]. It
is not enough to simply interpolate data points or pre-
dict new ones using advanced statistical methods such
as machine learning. Thus, even though such meth-
ods can approximate complex and expensive compu-
tations, naı̈ve applications of neural networks cannot
be agents of understanding. Scientific understanding
requires more than mere calculation. To illustrate this
point even further, let us consider one concrete exam-
ple in quantum physics from the literature: A com-
putational method solved an open question about the
generation of important resource states for quantum
computing. Then it extracted the conceptual core of
the solution in the form of a new quantum interfer-
ence effect in such a fashion that human scientists
can both understand the results and apply the ac-
quired understanding in different contexts [49]. Even

if the computer itself was able to apply the conceptual
core to other situations, it would not be a priori clear
whether the computer truly acquired scientific under-
standing. What is still missing is an explanation of
the discovered technique in the context of a scientific
theory. In this particular example, the android and
the human scientist would need to recognize the un-
derlying quantum interference in the context of the
theory of quantum physics. Thus, we can propose the
first sufficient condition for agents of understanding:
Condition for Scientific Understanding I:
An android gained scientific understanding if it
can recognize qualitatively characteristic conse-
quences of a theory without performing exact
computations and use them in a new context.
This condition closely follows the ideas of de Regt
and Dieks [13]. Let us go one step further and imagine
that there is an android capable of explaining discov-
eries in the context of scientific theories. How could
human scientists recognize that the machine acquired
new scientific understanding? We argue that human
scientists would do it in the exact same way they can
recognize that other human scientists acquired new
scientific understanding. That is, let the other human
scientists transfer the newly acquired understanding
to themselves. This allows us to propose the second
sufficient condition for agents of understanding:
Condition for Scientific Understanding II:
An android gained scientific understanding if it
can transfer its understanding to a human ex-
pert.
We argue that one can only recognize indirectly
whether a computer (or human) has gained scientific
understanding. Therefore, finally, we propose a test
in the spirit of the Turing test [41] or the Feigenbaum
test[109] (or adaptations thereof in the natural sci-
ences such as the Chemical Turing Test or the Feyn-
man Test [4]):
The Scientific Understanding Test:
A human (the student) interacts with a
teacher, either a human or an android sci-
entist. The teacher’s goal is to explain a sci-
entific theory and its qualitative, characteris-
tic consequences to the student. Another hu-
man (the referee) tests both the student and
the teacher independently3 . If the referee can-
not distinguish between the qualities of their
non-trivial explanations in various contexts,
we argue that the teacher has scientific un-
derstanding.
3 Inprinciple, there is no reason for the student or the referee not
to be androids. However, to keep the test as simple as possible,
we want to keep the number of possible variations small.
9
This implies that humans need to understand the
new concepts that androids devised. If a machine
truly understands something, it will be able to ex-
plain it and transfer the understanding to someone
else.4 We believe that this should always be possible,
even if the understanding is far beyond what human
experts know at this point. We envision that com-
puters will use advanced human-computer interaction
techniques together with the tools we described for
(future) computational microscopes.
Additionally, scientific discussions between a hu-
man and a computer could be realized using advanced
queries in natural language processing tools such as
BIRD [80] or GPT-3 [81]. That way, the scientist
could probe the computer with scientific questions.
Suppose the scientist gains new scientific understand-
ing by communicating with the algorithm, as judged
by our scientific understanding test. In that case, they
can confirm that the computer truly acquired under-
standing.5 We are optimistic that more efforts will
be directed at developing the necessary technologies,
which will lead to ever more convincing demonstra-
tions of android scientists acting as true agents of un-
derstanding in the future.
V. CONCLUSION
Undoubtedly, advanced computational methods in
general and artificial intelligence specifically will fur-
ther revolutionize how scientists investigate the se-
crets of our world. We outline how these new methods
can directly contribute to one of the main aims of sci-
ence, namely acquiring new scientific understanding.
We suspect that significant future progress in the use
of androids to acquire scientific understanding will re-
quire multidisciplinary collaborations between natu-
ral scientists, computer scientists and philosophers of
science. Thus, we firmly believe that these research
efforts can – within our lifetimes – transform androids
into true agents of understanding that will directly
contribute to one of the most essential aims of sci-
ence, namely Scientific Understanding.
ACKNOWLEDGEMENTS
The authors thank Anastassia Alexandrova, Rom-
mie Amaro, Curtis Berlinguette, Lillian Chong,
Gerardo Cisneros, Andy Cooper, Graeme Day,
4 We leave aside the question whether the explanation of the
android is true or false. It has been argued that also false
theories can lead to genuine understanding [110].
5 We would like to point out that our test, like the ones originated
by Turing and Feigenbaum, are not clear-cut, leaving room for
situations that do not allow a clear judgement.

Francois-Xavier Coudert, Lee Cronin, Elisa Fadda,
Rafael Gomez-Bombarelli, Leticia Gonzalez, Johannes
Hachmann, Roald Hoffmann, Jan Halborg Jensen,
Erin R. Johnson, Lynn Kamerlin, Heather J. Ku-
lik, Jean-Paul Malrieu, Anat Milo, Frank Noe,
Jens Kehlet Nørskov, Artem Oganov, Juan Perez-
Mercader, Chris Pickard, Markus Reiher, Jean-Louis
Reymond, Dennis Salahub, Stefano Sanvito, Franziska
Schoenebeck, Ilja Siepmann, Alex Sodt, Isaac Tam-
blyn, Donald Truhlar, Alexandre Tkatchenko, Koji
Tsuda, Alexandre Varnek, Tejs Vegge, Anatole von
[1] Lenka Zdeborová, “New tool in the box,” Nature
Physics 13, 420–421 (2017).
[2] Thomas Fösel, Petru Tighineanu, Talitha Weiss,
and Florian Marquardt, “Reinforcement learning
with neural networks for quantum feedback,” Phys-
ical Review X 8, 031084 (2018).
[3] Alexey A Melnikov, Hendrik Poulsen Nautrup,
Mario Krenn, Vedran Dunjko, Markus Tiersch, An-
ton Zeilinger, and Hans J Briegel, “Active learning
machine learns to create new quantum experiments,”
Proceedings of the National Academy of Sciences
115, 1221–1226 (2018).
[4] Alán Aspuru-Guzik, Roland Lindh, and Markus
Reiher, “The matter simulation (r)evolution,” ACS
central science 4, 144–152 (2018).
[5] Roald Hoffmann and Jean-Paul Malrieu, “Simula-
tion vs. understanding: A tension, in quantum chem-
istry and beyond. part a. stage setting,” Angewandte
Chemie 132, 12690–12710 (2020).
[6] Roald Hoffmann and Jean-Paul Malrieu, “Simula-
tion vs. understanding: A tension, in quantum chem-
istry and beyond. part b. the march of simulation, for
better or worse,” Angewandte Chemie International
Edition 59, 13156–13178 (2020).
[7] Roald Hoffmann and Jean-Paul Malrieu, “Simula-
tion vs. understanding: A tension, in quantum chem-
istry and beyond. part c. toward consilience,” Ange-
wandte Chemie International Edition 59, 13694–
13710 (2020).
[8] Gary Marcus, “The next decade in ai: four
steps towards robust artificial intelligence,”
arXiv:2002.06177 (2020).
[9] Jesse Thaler, “Designing an ai physicist,” CERN
Courier 9–10, 49 (2021).
[10] Angela Potochnik, “The diverse aims of science,”
Studies in History and Philosophy of Science Part
A 53, 71–80 (2015).
[11] Angela Potochnik, Idealization and the Aims of Sci-
ence (University of Chicago Press, 2017).
[12] Henk W De Regt, Understanding scientific under-
standing (Oxford University Press, 2017).
[13] Henk W De Regt and Dennis Dieks, “A contextual
approach to scientific understanding,” Synthese 144,
137–170 (2005).
[14] Henk W De Regt, “Understanding, values, and the
aims of science,” Philosophy of Science 87, 921–932
(2020).
10
Lilienfeld and Eva Zurek for answering our questions
on understanding, Xuemei Gu for Figure 2 and 3,
and Nora Tischler and Robert Fickler for valueable
comments on the manuscript. A.A.-G. and his group
acknowledge generous support from the Canada 150
Research Chairs Program, the University of Toronto,
and Anders G. Frøseth. M.K. acknowledges support
from the FWF (Austrian Science Fund) via the Erwin
Schrödinger fellowship No. J4309. R.P. acknowledges
funding through a Postdoc.Mobility fellowship by the
Swiss National Science Foundation (SNSF, Project
No. 191127).
[15] Carl G Hempel, Aspects of scientific explanation
(Free Press New York, 1965).
[16] Erwin Schrödinger, ’Nature and the Greeks’ and ’Sci-
ence and Humanism’ (Cambridge University Press,
1996).
[17] Henk W De Regt, “Visualization as a tool for un-
derstanding,” Perspectives on science 22, 377–396
(2014).
[18] Michael Friedman, “Explanation and scientific un-
derstanding,” The Journal of Philosophy 71, 5–19
(1974).
[19] Philip Kitcher, “Explanatory unification,” Philoso-
phy of science 48, 507–531 (1981).
[20] Werner Heisenberg, “über den anschaulichen inhalt
der quantentheoretischen kinematik und mechanik,”
Zeitschrift für Physik 43, 172–198 (1927).
[21] Raban Iten, Tony Metger, Henrik Wilming, Lı́dia
Del Rio, and Renato Renner, “Discovering physi-
cal concepts with neural networks,” Physical Review
Letters 124, 010508 (2020).
[22] Alireza Seif, Mohammad Hafezi, and Christopher
Jarzynski, “Machine learning the thermodynamic ar-
row of time,” Nature Physics 17, 105–113 (2021).
[23] Silviu-Marian Udrescu and Max Tegmark, “Ai feyn-
man: A physics-inspired method for symbolic regres-
sion,” Science Advances 6, eaay2631 (2020).
[24] John Jumper, Richard Evans, Alexander Pritzel,
Tim Green, Michael Figurnov, Olaf Ronneberger,
Kathryn Tunyasuvunakool, Russ Bates, Augustin
Žı́dek, Anna Potapenko, et al., “Highly accurate
protein structure prediction with alphafold,” Nature
596, 583–589 (2021).
[25] Kathryn Tunyasuvunakool, Jonas Adler, Zachary
Wu, Tim Green, Michal Zielinski, Augustin Žı́dek,
Alex Bridgland, Andrew Cowie, Clemens Meyer,
Agata Laydon, et al., “Highly accurate protein struc-
ture prediction for the human proteome,” Nature
596, 590–596 (2021).
[26] Lorenzo Casalino, Zied Gaieb, Jory A Goldsmith,
Christy K Hjorth, Abigail C Dommer, Aoife M Har-
bison, Carl A Fogarty, Emilia P Barros, Bryn C Tay-
lor, Jason S McLellan, et al., “Beyond shielding: the
roles of glycans in the sars-cov-2 spike protein,” ACS
Central Science 6, 1722–1734 (2020).
[27] Carl A Fogarty, Aoife M Harbison, Amy R Dug-
dale, and Elisa Fadda, “How and why plants and
human n-glycans are different: Insight from molecu-

lar dynamics into the ”glycoblocks” architecture of
complex carbohydrates,” Beilstein journal of organic
chemistry 16, 2046–2056 (2020).
[28] Pascal Friederich, Florian Häse, Jonny Proppe, and
Alán Aspuru-Guzik, “Machine-learned potentials for
next-generation matter simulations,” Nature Mate-
rials 20, 750–761 (2021).
[29] Sylvain Gigan, Florent Krzakala, Laurent Daudet,
and Igor Carron, “Artificial intelligence: From elec-
tronics to optics,” Photoniques 104, 49–52 (2020).
[30] Xingyuan Xu, Mengxi Tan, Bill Corcoran, Jiayang
Wu, Andreas Boes, Thach G Nguyen, Sai T Chu,
Brent E Little, Damien G Hicks, Roberto Moran-
dotti, et al., “11 tops photonic convolutional acceler-
ator for optical neural networks,” Nature 589, 44–51
(2021).
[31] Google AI Quantum et al., “Hartree-fock on a super-
conducting qubit quantum computer,” Science 369,
1084–1089 (2020).
[32] Jiehang Zhang, PW Hess, A Kyprianidis, P Becker,
A Lee, J Smith, G Pagano, I-D Potirniche, Andrew C
Potter, A Vishwanath, et al., “Observation of a dis-
crete time crystal,” Nature 543, 217–220 (2017).
[33] Christian Schweizer, Fabian Grusdt, Moritz Bern-
gruber, Luca Barbiero, Eugene Demler, Nathan
Goldman, Immanuel Bloch, and Monika Aidels-
burger, “Floquet approach to z2 lattice gauge theo-
ries with ultracold atoms in optical lattices,” Nature
Physics 15, 1168–1173 (2019).
[34] Esteban A Martinez, Christine A Muschik, Philipp
Schindler, Daniel Nigg, Alexander Erhard, Markus
Heyl, Philipp Hauke, Marcello Dalmonte, Thomas
Monz, Peter Zoller, et al., “Real-time dynamics of
lattice gauge theories with a few-qubit quantum
computer,” Nature 534, 516–519 (2016).
[35] Yudong Cao, Jonathan Romero, Jonathan P Ol-
son, Matthias Degroote, Peter D Johnson, Mária
Kieferová, Ian D Kivlichan, Tim Menke, Borja Per-
opadre, Nicolas PD Sawaya, et al., “Quantum chem-
istry in the age of quantum computing,” Chemical
reviews 119, 10856–10915 (2019).
[36] Christian Gross and Immanuel Bloch, “Quantum
simulations with ultracold atoms in optical lattices,”
Science 357, 995–1001 (2017).
[37] Michael O’Connor, Helen M Deeks, Edward Dawn,
Oussama Metatla, Anne Roudaut, Matthew Sutton,
Lisa May Thomas, Becca Rose Glowacki, Rebecca
Sage, Philip Tew, et al., “Sampling molecular confor-
mations and dynamics in a multiuser virtual reality
framework,” Science advances 4, eaat2731 (2018).
[38] Daniel Probst and Jean-Louis Reymond, “Exploring
drugbank in virtual reality chemical space,” Journal
of chemical information and modeling 58, 1731–1735
(2018).
[39] Jonas R Schmid, Moritz J Ernst, and Günther
Thiele, “Structural chemistry 2.0: Combining aug-
mented reality and 3d online models,” (2020).
[40] Davide Castelvecchi, “Using sound to explore events
of the universe,” Nature 597, 144 (2021).
[41] Alan M Turing, “Computing machinery and intelli-
gence,” Mind 50, 433–460 (1950).
[42] Joel Lehman, Jeff Clune, Dusan Misevic, Christoph
Adami, Lee Altenberg, Julie Beaulieu, Peter J Bent-
ley, Samuel Bernard, Guillaume Beslon, David M
11
Bryson, et al., “The surprising creativity of digital
evolution: A collection of anecdotes from the evolu-
tionary computation and artificial life research com-
munities,” Artificial life 26, 274–306 (2020).
[43] Chris J Pickard and RJ Needs, “Ab initio random
structure searching,” Journal of Physics: Condensed
Matter 23, 053201 (2011).
[44] Mario Krenn, Mehul Malik, Robert Fickler, Radek
Lapkiewicz, and Anton Zeilinger, “Automated
search for new quantum experiments,” Physical re-
view letters 116, 090405 (2016).
[45] Mario Krenn, Manuel Erhard, and Anton Zeilinger,
“Computer-inspired quantum experiments,” Nature
Reviews Physics 2, 649–661 (2020).
[46] Chris J Pickard and RJ Needs, “Highly compressed
ammonia forms an ionic crystal,” Nature materials
7, 775–779 (2008).
[47] Mario Krenn, Armin Hochrainer, Mayukh Lahiri,
and Anton Zeilinger, “Entanglement by path iden-
tity,” Physical review letters 118, 080401 (2017).
[48] Mario Krenn, Xuemei Gu, and Anton Zeilinger,
“Quantum experiments and graphs: Multiparty
states as coherent superpositions of perfect match-
ings,” Physical review letters 119, 240403 (2017).
[49] Mario Krenn, Jakob Kottmann, Nora Tischler,
and Alán Aspuru-Guzik, “Conceptual understand-
ing through efficient automated design of quantum
optical experiments,” Physical Review X 11, 031044
(2021).
[50] Pankaj Malhotra, Lovekesh Vig, Gautam Shroff,
and Puneet Agarwal, “Long short term memory net-
works for anomaly detection in time series,” Proceed-
ings on Atifical Neural Networks, Computational In-
telligence and Machine Learning (ESANN) 89, 89–
94 (2015).
[51] ATLAS Collaboration, “Dijet resonance search with
weak supervision using s= 13 tev p p collisions in the
atlas detector,” Physical review letters 125, 131801
(2020).
[52] CMS Collaboration, “Probing effective field theory
operators in the associated production of top quarks
with a z boson in multilepton final states at s=13
tev,” Journal of High Energy Physics 2021, 1–55
(2021).
[53] Sang Eon Park, Dylan Rankin, Silviu-Marian
Udrescu, Mikaeel Yunus, and Philip Harris, “Quasi
anomalous knowledge: searching for new physics
with embedded knowledge,” Journal of High Energy
Physics 2021, 1–26 (2021).
[54] Matthew D Schwartz, “Modern machine learning
and particle physics,” arXiv:2103.12226 (2021).
[55] Gregor Kasieczka, Benjamin Nachman, David Shih,
Oz Amram, Anders Andreassen, Kees Benk-
endorder, Blaz Bortolato, Gustaaf Broojimans, Flo-
rencia Canelli, Jack Collins, et al., “The lhc olympics
2020: a community challenge for anomaly detec-
tion in high energy physics,” Reports on Progress
in Physics (2021).
[56] Haizi Yu, Igor Mineyev, and Lav R Varshney,
“A group-theoretic approach to computational ab-
straction: Symmetry-driven hierarchical clustering,”
arXiv:1807.11167 (2018).
[57] Nima Dehmamy, Robin Walters, Yanchen Liu,
Dashun Wang, and Rose Yu, “Automatic symmetry

discovery with lie algebra convolutional network,”
Advances in Neural Information Processing Systems
34 (2021).
[58] AkshatKumar Nigam, Robert Pollice, Matthew FD
Hurley, Riley J Hickman, Matteo Aldeghi, Naruki
Yoshikawa, Seyone Chithrananda, Vincent A Voelz,
and Alán Aspuru-Guzik, “Assigning confidence to
molecular property prediction,” Expert Opinion on
Drug Discovery , 1–15 (2021).
[59] A Davies, P Velickovic, L Buesing, S Blackwell,
D Zheng, N Tomasev, R Tanburn, P Battaglia,
C Blundell, A Juhasz, et al., “Advancing mathe-
matics by guiding human intuition with ai,” Nature
(2021).
[60] Michael R Douglas, “Machine learning as a tool in
theoretical science,” Nature Reviews Physics , 1–2
(2022).
[61] Ross D King, Jem Rowland, Stephen G Oliver,
Michael Young, Wayne Aubrey, Emma Byrne, Maria
Liakata, Magdalena Markham, Pinar Pir, Larisa N
Soldatova, et al., “The automation of science,” Sci-
ence 324, 85–89 (2009).
[62] Anne-Catherine Bédard, Andrea Adamo, Kosi C
Aroh, M Grace Russell, Aaron A Bedermann,
Jeremy Torosian, Brian Yue, Klavs F Jensen, and
Timothy F Jamison, “Reconfigurable system for au-
tomated optimization of diverse chemical reactions,”
Science 361, 1220–1225 (2018).
[63] Sebastian Steiner, Jakob Wolf, Stefan Glatzel, Anna
Andreou, Jaroslaw M Granda, Graham Keenan,
Trevor Hinkley, Gerardo Aragon-Camarasa, Philip J
Kitson, Davide Angelone, et al., “Organic synthesis
in a modular robotic system driven by a chemical
programming language,” Science 363 (2019).
[64] Connor W Coley, Dale A Thomas, Justin AM Lum-
miss, Jonathan N Jaworski, Christopher P Breen,
Victor Schultz, Travis Hart, Joshua S Fishman, Luke
Rogers, Hanyu Gao, et al., “A robotic platform for
flow synthesis of organic compounds informed by ai
planning,” Science 365 (2019).
[65] Benjamin Burger, Phillip M Maffettone, Vladimir V
Gusev, Catherine M Aitchison, Yang Bai, Xiaoyan
Wang, Xiaobo Li, Ben M Alston, Buyi Li, Rob
Clowes, et al., “A mobile robotic chemist,” Nature
583, 237–241 (2020).
[66] Sourav Chatterjee, Mara Guidi, Peter H Seeberger,
and Kerry Gilmore, “Automated radial synthesis of
organic molecules,” Nature 579, 379–384 (2020).
[67] Jonathan Grizou, Laurie J Points, Abhishek Sharma,
and Leroy Cronin, “A curious formulation robot en-
ables the discovery of a novel protocell behavior,”
Science advances 6, eaay4237 (2020).
[68] Hyungil Moon, Dominic T Lennon, James Kirk-
patrick, Nina M van Esbroeck, Leon C Camenzind,
Liuqi Yu, Florian Vigneau, Dominik M Zumbühl,
G Andrew D Briggs, Michael A Osborne, et al., “Ma-
chine learning enables completely automatic tuning
of a quantum device faster than human experts,”
Nature communications 11, 1–10 (2020).
[69] Mogens Dalgaard, Felix Motzoi, Jens Jakob
Sørensen, and Jacob Sherson, “Global optimization
of quantum dynamics with alphazero deep explo-
ration,” npj Quantum Information 6, 1–9 (2020).
12
[70] Peder Larsen and Markus Von Ins, “The rate of
growth in scientific publication and the decline in
coverage provided by science citation index,” Scien-
tometrics 84, 575–603 (2010).
[71] James A Evans and Jacob G Foster, “Metaknowl-
edge,” Science 331, 721–725 (2011).
[72] Aaron Clauset, Daniel B Larremore, and Roberta
Sinatra, “Data-driven predictions in the science of
science,” Science 355, 477–480 (2017).
[73] Santo Fortunato, Carl T Bergstrom, Katy Börner,
James A Evans, Dirk Helbing, Staša Milojević,
Alexander M Petersen, Filippo Radicchi, Roberta
Sinatra, Brian Uzzi, et al., “Science of science,” Sci-
ence 359 (2018).
[74] Dashun Wang and Albert-László Barabási, The sci-
ence of science (Cambridge University Press, 2021).
[75] Vahe Tshitoyan, John Dagdelen, Leigh Weston,
Alexander Dunn, Ziqin Rong, Olga Kononova,
Kristin A Persson, Gerbrand Ceder, and Anubhav
Jain, “Unsupervised word embeddings capture latent
knowledge from materials science literature,” Nature
571, 95–98 (2019).
[76] Elsa A Olivetti, Jacqueline M Cole, Edward Kim,
Olga Kononova, Gerbrand Ceder, Thomas Yong-Jin
Han, and Anna M Hiszpanski, “Data-driven materi-
als research enabled by natural language processing
and information extraction,” Applied Physics Re-
views 7, 041317 (2020).
[77] Daniel Schwalbe-Koda, Zach Jensen, Elsa Olivetti,
and Rafael Gómez-Bombarelli, “Graph similarity
drives zeolite diffusionless transformations and in-
tergrowth,” Nature materials 18, 1177–1181 (2019).
[78] Andrey Rzhetsky, Jacob G Foster, Ian T Foster, and
James A Evans, “Choosing experiments to acceler-
ate collective discovery,” Proceedings of the National
Academy of Sciences 112, 14569–14574 (2015).
[79] Mario Krenn and Anton Zeilinger, “Predicting re-
search trends with semantic and neural networks
with an application in quantum physics,” Proceed-
ings of the National Academy of Sciences 117, 1910–
1916 (2020).
[80] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova, “Bert: Pre-training of deep bidi-
rectional transformers for language understanding,”
arXiv:1810.04805 (2018).
[81] Tom B Brown, Benjamin Mann, Nick Ryder,
Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,
Arvind Neelakantan, Pranav Shyam, Girish Sastry,
Amanda Askell, et al., “Language models are few-
shot learners,” arXiv:2005.14165 (2020).
[82] William L Hamilton, Rex Ying, and Jure Leskovec,
“Inductive representation learning on large graphs,”
in Proceedings of the 31st International Conference
on Neural Information Processing Systems (2017)
pp. 1025–1035.
[83] Grégoire Montavon, Wojciech Samek, and Klaus-
Robert Müller, “Methods for interpreting and un-
derstanding deep neural networks,” Digital Signal
Processing 73, 1–15 (2018).
[84] Wojciech Samek, Grégoire Montavon, Andrea
Vedaldi, Lars Kai Hansen, and Klaus-Robert
Müller, Explainable AI: interpreting, explaining and
visualizing deep learning, Vol. 11700 (Springer Na-
ture, 2019).

[85] Ribana Roscher, Bastian Bohn, Marco F Duarte,
and Jochen Garcke, “Explainable machine learning
for scientific insights and discoveries,” IEEE Access
8, 42200–42216 (2020).
[86] Scott M Lundberg, Gabriel Erion, Hugh Chen, Alex
DeGrave, Jordan M Prutkin, Bala Nair, Ronit Katz,
Jonathan Himmelfarb, Nisha Bansal, and Su-In Lee,
“From local explanations to global understanding
with explainable ai for trees,” Nature machine in-
telligence 2, 56–67 (2020).
[87] Aravindh Mahendran and Andrea Vedaldi, “Under-
standing deep image representations by inverting
them,” in Proceedings of the IEEE conference on
computer vision and pattern recognition (2015) pp.
5188–5196.
[88] Alexander Mordvintsev, Christopher Olah, and
Mike Tyka, “Inceptionism: Going deeper into neural
networks,” https://ai.googleblog.com/2015/06/
inceptionism-going-deeper-into-neural.html
(2015).
[89] Cynthia Shen, Mario Krenn, Sagi Eppel, and
Alan Aspuru-Guzik, “Deep molecular dreaming: In-
verse machine learning for de-novo molecular design
and interpretability with surjective representations,”
Machine Learning: Science and Technology (2021).
[90] Christopher P Burgess, Irina Higgins, Arka Pal, Loic
Matthey, Nick Watters, Guillaume Desjardins, and
Alexander Lerchner, “Understanding disentangling
in beta-vae,” arXiv:1804.03599 (2018).
[91] Pascal Friederich, Mario Krenn, Isaac Tamblyn, and
Alán Aspuru-Guzik, “Scientific intuition inspired by
machine learning-generated hypotheses,” Machine
Learning: Science and Technology 2, 025027 (2021).
[92] Daniel Flam-Shepherd, Tony Wu, Xuemei Gu, Alba
Cervera-Lierta, Mario Krenn, and Alan Aspuru-
Guzik, “Learning interpretable representations of en-
tanglement in quantum optics experiments using
deep generative models,” arXiv:2109.02490 (2021).
[93] Geemi P Wellawatte, Aditi Seshadri, and Andrew D
White, “Model agnostic generation of counterfac-
tual explanations for molecules,” Chemical Science
(2022).
[94] Thomas McGrath, Andrei Kapishnikov, Nenad
Tomašev, Adam Pearce, Demis Hassabis, Been Kim,
Ulrich Paquet, and Vladimir Kramnik, “Acquisition
of chess knowledge in alphazero,” arXiv:2111.09259
(2021).
[95] Michael Schmidt and Hod Lipson, “Distilling free-
form natural laws from experimental data,” Science
324, 81–85 (2009).
[96] Antonio A Gentile, Brian Flynn, Sebastian Knauer,
Nathan Wiebe, Stefano Paesani, Christopher E
Granade, John G Rarity, Raffaele Santagati, and
Anthony Laing, “Learning models of quantum sys-
tems from experiments,” Nature Physics 17, 837–843
(2021).
[97] Miles Cranmer, Alvaro Sanchez-Gonzalez, Pe-
ter Battaglia, Rui Xu, Kyle Cranmer, David
13
Spergel, and Shirley Ho, “Discovering symbolic
models from deep learning with inductive biases,”
arXiv:2006.11287 (2020).
[98] Kyle Cranmer, Johann Brehmer, and Gilles Louppe,
“The frontier of simulation-based inference,” Pro-
ceedings of the National Academy of Sciences 117,
30055–30062 (2020).
[99] Gal Raayoni, Shahar Gottlieb, Yahel Manor, George
Pisha, Yoav Harris, Uri Mendlovic, Doron Haviv,
Yaron Hadad, and Ido Kaminer, “Generating con-
jectures on fundamental constants with the ramanu-
jan machine,” Nature 590, 67–73 (2021).
[100] Adam Zsolt Wagner, “Constructions in combi-
natorics via neural networks,” arXiv:2104.14516
(2021).
[101] Pablo Lemos, Niall Jeffrey, Miles Cranmer, Shirley
Ho, and Peter Battaglia, “Rediscovering orbital me-
chanics with machine learning,” arXiv:2202.02306
(2022).
[102] Iyad Rahwan, Manuel Cebrian, Nick Obradovich,
Josh Bongard, Jean-François Bonnefon, Cynthia
Breazeal, Jacob W Crandall, Nicholas A Christakis,
Iain D Couzin, Matthew O Jackson, et al., “Machine
behaviour,” Nature 568, 477–486 (2019).
[103] Jürgen Schmidhuber, “Driven by compression
progress: A simple principle explains essential as-
pects of subjective beauty, novelty, surprise, interest-
ingness, attention, curiosity, creativity, art, science,
music, jokes,” Workshop on anticipatory behavior in
adaptive learning systems , 48–76 (2008).
[104] Deepak Pathak, Pulkit Agrawal, Alexei A Efros,
and Trevor Darrell, “Curiosity-driven exploration by
self-supervised prediction,” International Conference
on Machine Learning , 2778–2787 (2017).
[105] Luca A Thiede, Mario Krenn, AkshatKumar Nigam,
and Alan Aspuru-Guzik, “Curiosity in exploring
chemical space: Intrinsic rewards for deep molecular
reinforcement learning,” arXiv:2012.11293 (2020).
[106] Lav R Varshney, Nazneen Fatema Rajani, and
Richard Socher, “Explaining creative artifacts,”
arXiv:2010.07126 (2020).
[107] Lav R Varshney, Florian Pinel, Kush R Varsh-
ney, Debarun Bhattacharjya, Angela Schörgendor-
fer, and Y-M Chee, “A big data approach to compu-
tational creativity: The curious case of chef watson,”
IBM Journal of Research and Development 63, 7–1
(2019).
[108] Laurent Itti and Pierre Baldi, “Bayesian surprise at-
tracts human attention,” Vision research 49, 1295–
1306 (2009).
[109] Edward A Feigenbaum, “Some challenges and grand
challenges for computational intelligence,” Journal
of the ACM (JACM) 50, 32–40 (2003).
[110] Henk W De Regt and Victor Gijsbers, “How false
theories can yield genuine understanding,” Explain-
ing understanding: New perspectives from episte-
mology and philosophy of science , 50–75 (2017).